src/armnn/backends/NeonWorkloads/NeonDepthwiseConvolutionUint8Workload.cpp

   1 //
   2 // Copyright © 2017 Arm Ltd. All rights reserved.
   3 // See LICENSE file in the project root for full license information.
   4 //
   5
   6 #include "NeonDepthwiseConvolutionUint8Workload.hpp"
   7 #include "backends/NeonLayerSupport.hpp"
   8 #include "backends/CpuTensorHandle.hpp"
   9 #include "backends/ArmComputeTensorUtils.hpp"
  10
  11
  12 namespace armnn
  13 {
  14 using namespace armcomputetensorutils;
  15
  16 NeonDepthwiseConvolutionUint8Workload::NeonDepthwiseConvolutionUint8Workload(
  17     const DepthwiseConvolution2dQueueDescriptor& descriptor,
  18     const WorkloadInfo& info)
  19     : Uint8Workload<DepthwiseConvolution2dQueueDescriptor>(descriptor, info)
  20 {
  21     const TensorInfo& weightInfo = m_Data.m_Weight->GetTensorInfo();
  22
  23     m_KernelTensor = std::make_unique<arm_compute::Tensor>();
  24     BuildArmComputeTensor(*m_KernelTensor, weightInfo);
  25
  26     if (m_Data.m_Parameters.m_BiasEnabled)
  27     {
  28         m_BiasTensor = std::make_unique<arm_compute::Tensor>();
  29         BuildArmComputeTensor(*m_BiasTensor, m_Data.m_Bias->GetTensorInfo());
  30     }
  31
  32     arm_compute::PadStrideInfo padStrideInfo(m_Data.m_Parameters.m_StrideX,
  33                                              m_Data.m_Parameters.m_StrideY,
  34                                              m_Data.m_Parameters.m_PadLeft,
  35                                              m_Data.m_Parameters.m_PadRight,
  36                                              m_Data.m_Parameters.m_PadTop,
  37                                              m_Data.m_Parameters.m_PadBottom,
  38                                              arm_compute::DimensionRoundingType::FLOOR);
  39
  40     m_Data.ValidateInputsOutputs("NeonDepthwiseConvolutionUint8Workload", 1, 1);
  41
  42     arm_compute::ITensor& input  = static_cast<INeonTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
  43     arm_compute::ITensor& output = static_cast<INeonTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
  44
  45     bool use3x3Optimisation = weightInfo.GetShape()[3] == 3 && weightInfo.GetShape()[2] == 3;
  46     if (use3x3Optimisation)
  47     {
  48         m_pDepthwiseConvolutionLayer = std::make_unique<arm_compute::NEDepthwiseConvolutionLayer3x3>();
  49         static_cast<arm_compute::NEDepthwiseConvolutionLayer3x3*>(
  50             m_pDepthwiseConvolutionLayer.get())->configure(&input,
  51                                                            m_KernelTensor.get(),
  52                                                            m_BiasTensor.get(),
  53                                                            &output,
  54                                                            padStrideInfo);
  55     }
  56     else
  57     {
  58         m_pDepthwiseConvolutionLayer = std::make_unique<arm_compute::NEDepthwiseConvolutionLayer>();
  59         static_cast<arm_compute::NEDepthwiseConvolutionLayer*>(
  60             m_pDepthwiseConvolutionLayer.get())->configure(&input,
  61                                                            m_KernelTensor.get(),
  62                                                            m_BiasTensor.get(),
  63                                                            &output,
  64                                                            padStrideInfo);
  65     }
  66
  67     BOOST_ASSERT(m_pDepthwiseConvolutionLayer);
  68
  69     InitialiseArmComputeTensorData(*m_KernelTensor, m_Data.m_Weight->GetConstTensor<uint8_t>());
  70
  71     if (m_BiasTensor)
  72     {
  73         InitialiseArmComputeTensorData(*m_BiasTensor, m_Data.m_Bias->GetConstTensor<int32_t>());
  74     }
  75
  76     m_pDepthwiseConvolutionLayer->prepare();
  77     FreeUnusedTensors();
  78 }
  79
  80 void NeonDepthwiseConvolutionUint8Workload::Execute() const
  81 {
  82     ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonDepthwiseConvolutionUint8Workload_Execute");
  83     BOOST_ASSERT(m_pDepthwiseConvolutionLayer);
  84
  85     m_pDepthwiseConvolutionLayer->run();
  86 }
  87
  88 void NeonDepthwiseConvolutionUint8Workload::FreeUnusedTensors()
  89 {
  90     FreeTensorIfUnused(m_KernelTensor);
  91     FreeTensorIfUnused(m_BiasTensor);
  92 }
  93
  94 } //namespace armnn